Detection of a physical chess board state via Artificial Neural Networks

Sammanfattning

Translating the state of a physical chess board into a machine readable format is a problem that requires techniques within digital image processing and artificial neural networks. A solution to this problem has potential applications in chess analysis, training tools, and record keeping. This project aims to develop a system that converts an image of a chess board into its corresponding Forsyth-Edwards Notation (FEN). The system utilizes image processing methods to detect the chess board and extract the individual squares, and a convolutional neural network model for piece detection. The piece detection model achieved the precision and recall scores of 0.865 and 0.872 respectively, demonstrating a high degree of reliability and accuracy. However, the results of testing the complete system on a set of eight custom images separate from the dataset, revealed low overall performance, primarily due to classification errors made by the piece detection component. The system has met its initial goal of translating images to notations, albeit with limited accuracy, indicating possibilities for improvement in future iterations. One limitation is that the current program requires manual parameter adjustments during the grid detection process. Future improvements could aim to automate this step. Additionally, future versions of the system could focus on improving robustness, allowing the system to effectively work under varied lighting conditions, image qualities, and even support real-time detection.